The $856 Billion Factory Behind the Chatbot

Machines & Money - We follow the money through the machines that power modern life.

OpenAI's plan for the rest of the decade contains two numbers that seem like they can't both be true.

First, revenue rises from $36 billion this year to $350 billion in 2030, nearly tenfold.

Second, over the same 2026–2030 stretch, the company expects negative free cash flow of $278 billion. In plain terms, it spends $278 billion more than it takes in.

A business growing that fast is supposed to be swimming in cash. So where does it all go?

These figures come from a company presentation seen by the Financial Times, not audited financial statements. OpenAI has not published the projections, and a company finance executive has publicly disputed the reported figures. So treat them as a reported snapshot of OpenAI's planning, not financial guidance. Even as a plan, they reveal a lot about what it takes to make AI.

AI feels like software. Scaling it looks more like building a utility.

Asking ChatGPT a question feels weightless. Producing the answer is not.

Every response runs on specialized chips packed into racks, wired together with high-speed networking, and cooled by liquid plumbing. The racks sit inside buildings that need their own substations. Before any of it switches on, someone has to secure land, a power contract, and a place in line for the grid.

According to the FT, OpenAI expects to spend about $856 billion on computing power and infrastructure through 2030, its largest expense category.

What does that buy? Estimates for a gigawatt of AI capacity, chips included, range from about $35 billion (Bernstein) to Nvidia's own figure of $50–60 billion. Very roughly, $856 billion is on the order of 15 to 25 gigawatts.

Much of that isn't OpenAI buying buildings. It pays cloud partners like Microsoft and Oracle, who build the halls and bill for the capacity. The bill still lands in the same place.

AI feels like software. Scaling it looks more like building a utility.

Five years of revenue still doesn't cover the build.

Here is the strange part.

OpenAI projects roughly $840 billion in cumulative revenue through 2030. Over roughly the same period, it expects to spend about $856 billion on compute and infrastructure.

Those numbers aren't directly comparable dollar-for-dollar. One is revenue; the other is a spending category, and the timing and accounting don't line up perfectly. But together they show the scale of the problem: even extraordinary revenue growth may not be enough to finance the infrastructure required to produce it.

OpenAI still projects $278 billion in negative free cash flow through 2030.

The deeper problem is timing. The capacity has to exist before it can generate revenue. Chips get secured, data centers get built, and power gets contracted before the customers they enable have paid for them.

Growth can be real and still not fund the factory. The factory comes first.

Someone has to carry the gap.

So who pays in the meantime?

Investors do, first. OpenAI raised $122 billion in March at an $852 billion valuation, and the FT reports it's on track to exhaust that cash by 2028.

That makes the next financing part of the machine. Investors have discussed a new round valuing the company at $1.2 trillion or more. Public markets are the other route. OpenAI has confidentially submitted IPO paperwork to the SEC but hasn't set a timeline.

The gap doesn't sit only on OpenAI's books. The FT also reported that the future revenue expected by Nvidia, Oracle and SoftBank's data-center business depends heavily on OpenAI contracts. Those suppliers build and borrow against the promise that OpenAI keeps paying. The factory is financed in layers, and every layer is betting the revenue arrives on schedule.

There is some good news in the plan. The projected burn is smaller than it was in May, when the company expected $305 billion. The numbers are moving in the right direction. The shape hasn't changed.

The money moves before the verdict.

Whether OpenAI's bet pays off is a question for 2030. The spending is happening now, and much of it goes to things that have nothing to do with which AI model wins.

Chips are the obvious beneficiary. Less obvious is everything underneath the rack.

Every new AI campus needs transformers, switchgear, breakers, cooling, and a connection to the grid. Lead times for some high-voltage transformers have stretched to multiple years, up from around a year in 2020 and 2021. Wood Mackenzie estimates data centers could make up as much as 40% of the electrical-equipment market in accelerated scenarios, up from just under 2% in 2020.

Those suppliers don't need ChatGPT to beat its rivals. They don't need AI revenue to hit $350 billion. They need the building to continue.

That's the quiet logic of this moment. Eventually, the AI race will be judged by software margins. But the money is moving long before that verdict arrives.

For now, AI is being built in copper, steel, silicon and transformers. And the first money through the machine goes to whoever sells the factory its parts.